An image-based insulating film defect identification method and system

Abstract

The application discloses a kind of based on image's heat insulation film defect identification method and system, it is related to industrial vision detection technical field, including, establish heat insulation film polarization imaging environment and gather displacement synchronous original image, to original polarization image is carried out imaging correction and establishes heat insulation film physical coordinate image;To standardization heat insulation film image is carried out directionality texture analysis and filters speckle candidate area, gathers the good product image of each heat insulation film, establishes heat insulation film normal period reference using current batch good product image;To speckle candidate area is extended, extract one-dimensional gray signal of speckle candidate area and quantize period damage degree;Establish batch self-adapting elliptical decision boundary and to speckle candidate area is confirmed right, further calculate speckle visibility index and output classification alarm and roll position.The application realizes to the quantization of function particle period arrangement damage, enhances the determination adaptability under different batch conditions.

Description

Technical Field

[0001] This invention relates to the field of industrial visual inspection technology, and in particular to an image-based method and system for identifying defects in heat insulation films. Background Technology

[0002] Heat insulation films are widely used in architectural glass, automotive glass, and transparent heat insulation components. They typically use polyester polymer films as the substrate and form metal layers, ceramic functional particle layers, or other photothermal control layers through processes such as vacuum coating, magnetron sputtering, coating dispersion, or composite lamination. During continuous coating, stretching, curing, and winding, heat insulation films are prone to developing blemishes, light and dark bands, and uneven light transmission due to uneven dispersion of functional particles, local agglomeration or absence, coating thickness fluctuations, roller surface contamination, and tension changes. Early quality inspection mainly relied on manual visual inspection under specified lighting and observation distances. However, manual inspection is affected by the inspector's experience, fatigue level, and observation environment, making it difficult to meet the continuous inspection needs of high-speed roll material production. With the development of industrial machine vision technology, existing heat insulation film or transparent film inspection systems are gradually adopting area scan cameras or line scan cameras. In conjunction with backlighting, the grayscale, edges, color difference, and texture of the film surface are analyzed. To reduce specular reflection on the transparent film surface, some systems also use polarizers, diffuse light sources, or multi-angle illumination, and employ grayscale thresholding, grayscale co-occurrence matrix, Fourier transform, wavelet analysis, directional filtering, and machine learning models to identify periodic texture anomalies. However, heat insulation film defects have the characteristics of transparent medium imaging, functional particle texture, and continuous roll material movement being coupled together. Existing general surface visual inspection methods still have certain limitations. Under the action of stretching and coating, the functional particles of the heat insulation film usually form textures with certain periodicity and directionality along the production direction. Simply using two-dimensional statistical features is difficult to distinguish between the slow changes in normal texture and the destruction of local periodic structures. Using only the global Fourier spectrum is also easily affected by changes in production speed, differences in materials from different batches, and differences in film width position. Summary of the Invention

[0003] In view of the aforementioned existing problems, the present invention is proposed.

[0004] Therefore, the present invention provides an image-based method and system for identifying defects in heat insulation films, which solves the problem that existing technologies cannot simultaneously consider periodic and directional textures in defect detection.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides an image-based method for identifying defects in thermal insulation films, comprising, Establish a polarization imaging environment for the heat insulation film and acquire original images with displacement synchronization. Perform imaging correction on the original polarization images and establish a physical coordinate image of the heat insulation film. Directional texture analysis was performed on standardized heat insulation film images to screen candidate areas for patterns. Good product images of each type of heat insulation film were collected, and a normal cycle benchmark for heat insulation film was established using the good product images of the current batch. The candidate regions for blemishes are expanded, and the one-dimensional grayscale signals of the candidate regions are extracted and the degree of periodic damage is quantified. The degree of periodic disruption includes the amplitude of the main frequency maintenance and high-frequency noise characteristics; Establish batch adaptive elliptical decision boundaries and assign weights to candidate areas of the blemishes. For each assigned blemish area, refine the pixel boundaries of the blemish defects and calculate the physical area and visual deviation. Further calculate the blemish visibility index and output graded alarms and roll material location.

[0006] As a preferred embodiment of the image-based heat insulation film defect identification method of the present invention, the establishment of the heat insulation film polarization imaging environment and the acquisition of original images with displacement synchronization are specifically as follows: a high-definition line array camera is vertically installed above the heat insulation film production line, so that the pixel arrangement direction of the line array camera is consistent with the width direction of the heat insulation film. When the heat insulation film moves continuously along the production line, the line array camera acquires images line by line, and each acquired line is stitched together in sequence according to the movement order of the film body to form a two-dimensional original image that simultaneously covers the longitudinal and transverse positions of the heat insulation film. A uniform backlight is set on the other side of the heat insulation film corresponding to the line scan camera, so that the backlight passes through the heat insulation film and enters the line scan camera. A first polarizer is set on the light-emitting surface of the backlight, and a second polarizer is set in front of the camera lens. The transmission directions of the two polarizers are set to be approximately orthogonal, and the included angle between the two polarizers is calibrated. A rotary encoder is installed on the production line drive roller or guide roller corresponding to the line scan camera and connected to the line scan camera to define the physical size of the longitudinal pixels and the physical size of the transverse pixels of the heat insulation film.

[0007] As a preferred embodiment of the image-based heat insulation film defect identification method of the present invention, the step of performing imaging correction on the original polarization image and establishing the physical coordinate image of the heat insulation film specifically involves: turning off the backlight before formal production testing, keeping the gain, exposure time, and acquisition frequency of the line scan camera consistent with the formal testing state, continuously acquiring 100 lines of dark field images, calculating the average value of the dark field grayscale 100 times for each pixel position in the line scan camera to obtain the dark field reference for that pixel position, then turning on the backlight, placing a good quality heat insulation film that has been manually confirmed to be free of blemishes and has uniform light transmission at the testing position, continuously acquiring 100 lines of uniform field images, calculating the average grayscale for each line scan pixel position to obtain the uniform field response for that pixel position, and for each original pixel acquired during formal production, first subtracting the dark field reference corresponding to the same line scan position, and then dividing by the difference between the uniform field response and the dark field reference for that position to obtain the normalized corrected grayscale, setting the calculation result less than 0 to 0, and setting the calculation result greater than 1 to 1, so that the corrected grayscale is within the range of 0 to 1; By utilizing the grayscale difference between the membrane area and the background of the field of view, the left and right boundaries of the heat insulation membrane are found in each image row, and the background area outside the boundaries is excluded. The encoder's cumulative position when a roll of heat insulation film enters the inspection station is defined as the longitudinal starting point, and the lateral position corresponding to the left edge of the heat insulation film is defined as the lateral starting point. Based on the longitudinal and lateral starting points, the original image pixels are mapped to the heat insulation film coordinates to form a standardized heat insulation film image.

[0008] As a preferred embodiment of the image-based heat insulation film defect identification method of the present invention, the step of performing directional texture analysis on the standardized heat insulation film image and screening candidate areas of spots specifically involves: dividing the standardized heat insulation film image into overlapping rectangular image blocks, performing local gray-level normalization on each image block, calculating the median gray level of the image block, and then calculating the median absolute deviation of the gray level of each pixel relative to the median. The median absolute deviation is used to represent the local gray-level fluctuation scale of the image block. The locally normalized gray level is quantified into 32 levels, and gray-level co-occurrence matrices are established along the stretching direction and the width direction of the heat insulation film, respectively. Two pixel distances are used in each direction. For each direction and each pixel distance, four texture statistics, namely energy, entropy, contrast and correlation, are extracted from the gray-level co-occurrence matrix. Weighted fusion is performed on texture amounts of the same type at a distance of two pixels, and further weighted fusion is performed on texture amounts in the stretching direction and the film width direction; For each type of heat insulation film, good product image blocks are collected in advance, and the normal median and normal fluctuation range of its energy, entropy, contrast and correlation are statistically analyzed. For the image patch to be detected, the degree of decrease in energy relative to the good benchmark, the degree of increase in entropy relative to the good benchmark, the degree of increase in contrast, and the degree of decrease in correlation are calculated and a texture anomaly score is constructed. The 99th percentile of the texture anomaly score of the good image patch is determined as the candidate screening threshold. When the texture anomaly score of the image patch to be detected exceeds the candidate screening threshold, it is marked as a pattern candidate patch; otherwise, it is identified as a normal texture image patch. Candidate image patches that overlap spatially or are connected by eight neighboring regions are merged into the same pattern candidate region, and the vertical start and end positions, horizontal start and end positions, center position, coverage area, and maximum texture anomaly score of the candidate region are recorded.

[0009] As a preferred embodiment of the image-based heat insulation film defect identification method of the present invention, the step of establishing a normal cycle benchmark for the heat insulation film using the images of good products in the current batch specifically involves: after each batch of heat insulation film starts production and enters a stable operating state, selecting good product image windows that have been manually confirmed to be free of blemishes from the images of the current batch, with each good product image window covering 20 normal texture cycles along the stretching direction; for each good product image window, averaging the pixel grayscale at the same longitudinal position along the film width direction to obtain a one-dimensional grayscale signal that varies along the stretching direction of the heat insulation film. Linear detrending processing is performed on the one-dimensional grayscale signal. Hanning windows are applied to both ends of the detrended signal. The one-dimensional grayscale signal is converted from the spatial domain to the spatial frequency domain using fast Fourier transform. Local spectral peaks are found within the effective frequency range. The maximum spectral peak is determined. Among all local peaks with an amplitude not less than 60% of the maximum peak, the local peak with the lowest frequency is selected as the fundamental frequency of the good window. The median of the base frequencies corresponding to the 20 good product image windows is determined as the normal texture frequency of the current batch, and its reciprocal is determined as the normal average texture period.

[0010] As a preferred embodiment of the image-based heat insulation film defect identification method of the present invention, the step of extracting the one-dimensional grayscale signal of the candidate region of the blemish and quantifying the periodic damage degree is as follows: for each candidate region of the blemish, taking the center position of the candidate region as a reference, the spectrum analysis window is extended forward and backward along the stretching direction of the heat insulation film. The extended spectrum analysis window covers 20 normal texture cycles along the stretching direction. In the horizontal direction, based on the boundary of the candidate region, the width of the candidate region is extended to the left and right sides by one-quarter respectively. The pixel grayscale of the same vertical position in the horizontal range is averaged to obtain the one-dimensional grayscale signal of the candidate region. Linear detrending, robust scaling and Hanning window processing are performed sequentially on the one-dimensional grayscale signal of the candidate region. Fast Fourier Transform is performed on the processed one-dimensional signal. The normal main frequency band is defined according to the normal main frequency. The second harmonic band is set near twice the normal main frequency. The main frequency holding amplitude of the candidate region is calculated. The high-frequency diffusion energy and non-dominant frequency diffusion energy of the candidate region are further calculated. The high-frequency diffusion energy and non-dominant frequency diffusion energy are weighted and fused into high-frequency noise features. The dominant frequency amplitude and high-frequency noise features are combined to form two-dimensional spectral feature points. At the same time, the complete spectral curve, dominant frequency peak position, analysis window range and texture anomaly score of the candidate region are saved.

[0011] As a preferred embodiment of the image-based heat insulation film defect identification method of the present invention, the step of establishing batch adaptive elliptical decision boundaries and determining the weight of candidate regions for blemishes specifically involves: For each type of heat insulation film, a model-level normal spectrum feature library is established, and the normal spectrum feature center and normal spectrum feature covariance matrix are calculated based on the model-level normal spectrum feature library. The normal spectrum model at the model level is adaptively adjusted using the spectral characteristics of the current batch of good products. The adjusted normal feature center is obtained by weighting the historical feature center at the model level and the feature center of the current batch of good products. The adjusted covariance matrix is ​​shrunk and regularized. For the candidate region of the blemish, the elliptical distance is calculated based on the adjusted normal feature center and the shrunk covariance matrix. The corresponding elliptical distance is calculated using the model-level good product samples. The 99.5 percentile of the good product elliptical distance is determined as the blemish weighting threshold. When the elliptical distance of the candidate region is less than or equal to the blemish weighting threshold, it is judged as normal texture fluctuation. If the elliptical distance of the candidate region is greater than the blemish weighting threshold, it is judged as a real blemish defect region. For multiple confirmed regions that overlap spatially, are directly adjacent, or are spaced apart by less than one normal texture cycle, they are merged into the same pattern defect object. If the interval between two confirmed regions is greater than one normal texture cycle, and the texture and spectrum of the middle region have been restored to normal, they are retained as two independent pattern defects.

[0012] As a preferred embodiment of the image-based heat insulation film defect identification method of the present invention, the step of refining the pixel boundaries of the mottled defects and calculating the physical area and visual deviation degree specifically involves: for each confirmed mottled area, establishing a local background ring around it; when the number of effective background pixels is less than 60% of the total number of pixels in the background ring, gradually expanding the range of the background ring until a sufficient number of normal background pixels are obtained; if a sufficient number of background pixels still cannot be obtained after expansion, the defect is marked as an edge region defect and supplemented using the background parameters of the nearest normal image block. A local grayscale plane is fitted using grayscale pixels within the background ring band. The local grayscale plane is used to represent the normal grayscale that each position in the confirmation area should have when there are no spots. For each pixel in the confirmation area, the absolute difference between its actual corrected grayscale and the corresponding local background grayscale is calculated. Then, based on the residual between the actual grayscale in the background ring band and the fitted background grayscale, the background robust fluctuation scale is calculated. Pixels with a grayscale difference exceeding 3 times the background robust fluctuation scale are marked as initial spot pixels. Connectivity processing is performed on the initial blemish pixels, and the pixel region connected to the center of the confirmation region is determined as the final blemish mask. The number of pixels contained in the final blemish mask is counted and converted into the actual physical area. For each blemish pixel, the absolute difference between its actual gray level and the local background gray level is calculated, and then divided by the corresponding local background gray level to obtain the relative gray level deviation of the pixel. The average relative gray level deviation of all blemish pixels is calculated to obtain the average relative gray level deviation of the blemish defect. At the same time, the longitudinal length, transverse width, aspect ratio, contour perimeter, and distance from the left and right edges of the membrane are calculated for the blemish defect.

[0013] As a preferred embodiment of the image-based heat insulation film defect identification method of the present invention, the calculation of the mottle visibility index and the output of graded alarms and roll material location specifically includes: For each blemish defect, calculate the blemish visibility index; Obtain the acceptance threshold for each type of heat insulation film. Determine the defect level of the heat insulation film based on the pattern visibility index and the acceptance threshold. If the pattern visibility index is less than or equal to the acceptance threshold, it is judged as level 0, allowing texture fluctuations. No production alarm is triggered, and only the test data is saved. If the pattern visibility index is greater than the acceptance threshold but less than or equal to twice the acceptance threshold, it is judged as level 1, visually visible pattern. Mark the pattern position in the test interface, trigger an audio-visual prompt, and generate a defect local image and outline for the operator to review. When the pattern visibility index is greater than twice the acceptance threshold, it is judged as level 2, obviously unqualified pattern. A high-level alarm is output, and a coding mark, labeling, speed reduction, or shutdown signal is sent according to the production line control requirements. Based on the mapping relationship between the original image and the coordinates of the heat insulation film, the longitudinal distance from the start end of the roll of each blemish defect, the lateral distance from the left edge of the heat insulation film, and the longitudinal and lateral start and end positions of the defect are calculated. Defect data is output for each blemish. For the same type of blemish that appears continuously in the same roll of heat insulation film, trend statistics are performed according to the defect location and frequency of occurrence. When multiple level 2 blemishes appear continuously within a short longitudinal distance, or when the number of level 1 blemishes per unit length continues to increase, a process trend warning is output. After a roll of heat insulation film is completed, the normal areas and blemish defect areas that have been manually verified are saved to the sample library of the corresponding model.

[0014] Secondly, the present invention provides an image-based thermal insulation film defect identification system, comprising, The image processing module is used to establish the polarization imaging environment of the heat insulation film and acquire the original image with displacement synchronization, perform imaging correction on the original polarization image and establish the physical coordinate image of the heat insulation film. The benchmark analysis module is used to perform directional texture analysis on standardized heat insulation film images and screen candidate areas for patterns, collect good product images of each type of heat insulation film, and establish a normal cycle benchmark for heat insulation film using the good product images of the current batch. The candidate region quantization module is used to expand the mottled candidate region, extract the one-dimensional grayscale signal of the mottled candidate region, and quantify the degree of periodic damage. The pattern recognition output module is used to establish batch adaptive elliptical decision boundaries and assign weights to candidate pattern regions. For each assigned pattern region, the pixel boundaries of the pattern defect are refined and the physical area and visual deviation are calculated. The pattern visibility index is further calculated and the graded alarm and roll material location are output.

[0015] The beneficial effects of this invention are as follows: By using displacement-synchronous polarization acquisition and dark field / uniform field correction, the image rows stably correspond to the actual position of the membrane, suppressing reflection, uneven illumination, and pixel response differences; by using a directional gray-level co-occurrence matrix for initial screening, the number of normal areas entering spectral analysis is reduced, improving online processing efficiency; by batch main frequency calibration, main frequency amplitude preservation, and non-main frequency dispersion energy extraction, the destruction of the periodic arrangement of functional particles is quantified; by jointly constructing elliptical boundaries using historical model and current batch features, the adaptability of judgment under different batch conditions is enhanced; by refining defect boundaries and converting physical area through local background fitting, the accuracy of positioning and size evaluation is improved; by using visibility index, visual calibration threshold, and roll material coordinate output, the detection results can serve re-inspection, coding, slitting, and production alarms, thereby reducing excessive rejection caused by micro-texture fluctuations. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of the image-based heat insulation film defect identification method in Example 1.

[0018] Figure 2 This is a diagram showing the relationship between candidate screening for mottled patterns and periodic weighting.

[0019] Figure 3This is a structural diagram of the image-based heat insulation film defect identification system in Example 1. Detailed Implementation

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0021] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0022] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0023] Example 1, referring to Figures 1-3 This is the first embodiment of the present invention, which provides an image-based method for identifying defects in heat insulation films, including the following steps: S1. Establish the polarization imaging environment of the heat insulation film and acquire the original image with displacement synchronization. Perform imaging correction on the original polarization image and establish the physical coordinate image of the heat insulation film. S1.1 A high-definition line array camera is vertically installed above the heat insulation film production line, so that the pixel arrangement direction of the line array camera is consistent with the width direction of the heat insulation film. When the heat insulation film moves continuously along the production line, the line array camera collects images line by line. Each collected line is spliced ​​in sequence according to the movement order of the film body to form a two-dimensional original image that simultaneously covers the longitudinal and transverse positions of the heat insulation film. A uniform backlight is set on the other side of the heat insulation film corresponding to the line scan camera, so that the backlight passes through the heat insulation film and enters the line scan camera. A first polarizer is set on the light-emitting surface of the backlight, and a second polarizer is set in front of the camera lens. The transmission directions of the two polarizers are set to be approximately orthogonal. The included angle between the two polarizers is calibrated. The cross-polarization imaging structure is adopted because the specular reflection light generated on the surface of the heat insulation film usually maintains its original polarization direction and can be weakened by the second polarizer in front of the lens. The light that passes through the heat insulation film and is scattered by metal particles, ceramic particles or film microstructure changes its polarization state and can still partially pass through the second polarizer to enter the camera. This imaging structure can reduce the interference of film surface reflection and highlight the distribution texture of functional particles inside the heat insulation film. Specifically, the calibration process for the angle between polarizers involves first setting the angle between the two polarizers to a theoretical orthogonal angle of 90°, and then rotating the second polarizer in front of the lens gradually within the range of 85° to 95°. For every 1° rotation, an image of a good quality heat insulation film without obvious blemishes is acquired, and the 99th percentile grayscale value of each image is calculated. The angle corresponding to the lowest 99th percentile grayscale value is determined as the working polarization angle. The 99th percentile grayscale value is used for calibration instead of the maximum grayscale value because the maximum grayscale value is easily affected by camera dead pixels, isolated reflective points, or dust spots, while the 99th percentile grayscale value can reflect the residual reflectivity of the overall bright areas of the image. A rotary encoder is installed on the drive roller or guide roller of the production line corresponding to the line scan camera and connected to the line scan camera. Each time the encoder detects that the heat insulation film has moved a set physical distance, it sends an image acquisition trigger signal to the line scan camera. The physical size of the vertical pixels of the heat insulation film is defined as... The horizontal pixel physical size is Then, for every movement of the heat insulation film Acquire one line of image per millimeter; The physical size of the vertical pixel is and horizontal pixel physical size The minimum detectable blemish size is determined to ensure that the minimum detectable blemish covers at least 10 pixels in any direction. For example, if the minimum blemish width to be detected is 0.5 mm, the physical size of the vertical pixels can be determined as follows: and horizontal pixel physical size Set to 0.05 mm per pixel; The camera exposure time is set according to the maximum operating speed of the production line, so that the longitudinal movement distance of the heat insulation film during the exposure does not exceed 0.3 times the physical size of the longitudinal pixel. This can limit motion blur to the sub-pixel range. The specific exposure time can be determined by dividing the maximum speed of the production line by the physical size of the longitudinal pixel. Under the premise of meeting the motion blur requirements, the imaging grayscale is guaranteed by the backlight brightness.

[0024] S1.2 Before formal production testing, turn off the backlight, keep the gain, exposure time, and acquisition frequency of the line scan camera consistent with the formal testing state, and continuously acquire 100 lines of dark field images. For each pixel position in the line scan camera, calculate the average value of the dark field grayscale 100 times to obtain the dark field reference for that pixel position. The dark field reference is used to reflect the dark current, readout bias, and fixed mode noise of the camera under conditions of no effective incident light. Then turn on the backlight, place a good quality heat insulation film that has been manually confirmed to be free of spots and have uniform light transmission at the testing position, and continuously acquire 100 lines of dark field images. For each linear array pixel position in the uniform field image, the average gray level is calculated to obtain the uniform field response at that pixel position. The uniform field response reflects the lateral brightness difference of the backlight, lens vignetting, and the sensitivity difference of the linear array pixels. For each original pixel acquired during the formal production process, the dark field reference corresponding to the same linear array position is subtracted first, and then divided by the difference between the uniform field response and the dark field reference at that position to obtain the normalized corrected gray level. The calculation result less than 0 is set to 0, and the calculation result greater than 1 is set to 1, so that the corrected gray level is in the range of 0 to 1. By utilizing the grayscale difference between the membrane area and the background of the field of view, the left and right boundaries of the heat insulation membrane are found in each image row, and the background area outside the boundaries is excluded. If the heat insulation film shifts laterally during operation, the effective detection range of the film is updated according to the edge position change of the continuous image lines, so that the image blocks always act on the real film area, rather than a fixed camera pixel range. The encoder's cumulative position when a roll of heat insulation film enters the inspection station is defined as the longitudinal starting point, and the lateral position corresponding to the left edge of the heat insulation film is defined as the lateral starting point. Based on the longitudinal and lateral starting points, the original image pixels are mapped to the heat insulation film coordinates to form a standardized heat insulation film image. When forming the standardized heat insulation film image, the product model, production batch, roll number, acquisition time, encoder position, polarization angle, camera exposure time, and pixel physical size need to be written into the corresponding image data for subsequent defect traceability.

[0025] S2. Perform directional texture analysis on standardized heat insulation film images and screen candidate areas for patterns. Collect good product images of each type of heat insulation film and establish a normal cycle benchmark for heat insulation film using the good product images of the current batch. S2.1 Divide the standardized heat insulation film image into overlapping rectangular image blocks. The vertical and horizontal physical dimensions of each image block are at least four times the smallest detectable stain size, ensuring that each image block contains both defective areas and a certain range of surrounding normal textures as a reference. The vertical and horizontal movement step of each image block is half the corresponding side length. Perform local gray-level normalization on each image block, calculate the median gray level of the image block, and then calculate the median absolute deviation of each pixel's gray level relative to this median. Use the median absolute deviation to represent the local gray-level fluctuation scale of the image block. Quantize the locally normalized gray levels into 32 levels. The original 12-bit image contains 4096 gray levels. If 40 levels are directly established... A 96×4096 gray-level co-occurrence matrix would result in an extremely sparse matrix and excessive computational cost. Quantization to 32 levels can preserve the main gray-level changes caused by the spots and ensure that various gray-level combinations have sufficient statistical frequency. Gray-level co-occurrence matrices are established along the stretching direction and the width direction of the heat insulation film, respectively. Two pixel distances are used in each direction. The first pixel distance is set to 1 pixel to reflect the edges of adjacent particles and fine-grained textures. The second pixel distance is set to the number of pixels corresponding to half the width of the smallest spot to be detected, to reflect the medium-scale textures formed by particle aggregation, absence, or local density changes. For each direction and each pixel distance, four texture statistics, namely energy, entropy, contrast, and correlation, are extracted from the gray-level co-occurrence matrix. Energy is used to indicate whether the grayscale combination is concentrated. Normal heat insulation films have relatively stable functional particle arrangements and typically high energy, while patterns disperse the grayscale combination and typically decrease the energy. Entropy is used to represent the degree of disorder in grayscale combinations. Particle clustering, particle absence, or uneven distribution increase local randomness, thus increasing entropy. Contrast is used to represent the degree of gray-level jump between adjacent or spaced pixels. Particle cluster boundaries, empty area boundaries, or mottled edges usually increase contrast. Correlation is used to indicate whether the gray levels of pixels separated by a certain distance maintain a stable relationship. Normal periodic textures have high correlation, while correlation usually decreases after the pattern breaks the continuity of the arrangement. The texture quantities of the same type between two pixels are weighted and fused. The adjacent distance weight is preferably 0.4, and the mesoscale distance weight is preferably 0.6. The mesoscale distance weight is higher because the present invention mainly identifies the patterns formed by functional particle aggregation, absence, or uneven density, rather than the small changes produced by the edge of a single particle. If the adjacent distance weight is too high, the detection results are easily affected by particle edge noise. If the mesoscale distance weight is too high, it may reduce the sensitivity to fine patterns. The texture quantities in the stretching direction and the film width direction are further weighted and fused. The preferred fusion weights are 0.7 in the stretching direction and 0.3 in the film width direction. The stretching direction weight is higher because during the stretching, coating, and curing process of the heat insulation film, the internal functional particles usually form obvious extended textures along the film's direction. Patterns will preferentially disrupt the continuity in this direction. Retaining a film width direction weight of 0.3 can identify aggregation areas that diffuse laterally. For each type of heat insulation film, no less than 500 good image blocks that have been manually confirmed to be normal are collected in advance, and their normal median values ​​and normal fluctuation ranges of energy, entropy, contrast and correlation are statistically analyzed. For each image patch to be detected, the decrease in energy relative to the benchmark, the increase in entropy relative to the benchmark, the increase in contrast, and the decrease in correlation are calculated to construct a texture anomaly score. The texture anomaly score is obtained by weighting, with the preferred weight values ​​being 0.20, 0.35, 0.25, and 0.20. The entropy weight is set to 0.35 because disordered particle distribution usually manifests first as an increase in grayscale combination disorder. The contrast weight is set to 0.25 to highlight particle cluster boundaries and particle void boundaries. The energy and correlation weights are set to 0.20 to help evaluate the degree of texture concentration and arrangement continuity. The 99th percentile of the texture anomaly score of the benchmark image patch is determined as the candidate screening threshold. When the texture anomaly score of the image patch to be detected exceeds the candidate screening threshold, it is marked as a pattern candidate patch; otherwise, it is considered a normal texture image patch. Candidate image patches that overlap spatially or are connected by eight neighboring regions are merged into the same pattern candidate region, and the vertical start and end positions, horizontal start and end positions, center position, coverage area, and maximum texture anomaly score of the candidate region are recorded.

[0026] S2.2 After each batch of heat insulation film starts production and enters a stable operating state, select no less than 20 good product image windows from the current batch images that have been manually confirmed to be free of blemishes. Good product windows must not contain film edges, roll joints, wrinkles, contamination, damage or obvious lighting abnormalities. Each good product image window covers 20 normal texture cycles along the stretching direction. The use of at least 20 cycles is to ensure that the Fourier analysis has sufficient frequency resolution and to reduce the impact of single local texture changes on the judgment of normal main frequency. For each good product image window, the pixel grayscale at the same longitudinal position along the film width direction is averaged to obtain a one-dimensional grayscale signal that varies along the stretching direction of the heat insulation film. At the initial moment, the historical average normal cycle of this model is read as the initial cycle. The first good product window is extracted at 20 times the initial cycle. The current batch main frequency and cycle are calculated based on the first window. The window length is readjusted according to the current batch cycle and the calculation is performed again. The result of the second calculation is used as the final normal main frequency of the current batch. This one-dimensional signal reflects the overall light transmittance or light blocking degree of functional particles at different longitudinal positions of the heat insulation film. The reason for using lateral averaging instead of extracting only single-row pixel signals is that single-row pixel signals are easily affected by camera dead pixels, isolated particles, dust and local micro scratches. Lateral averaging can enhance the periodic texture that coexists within a certain width range, while reducing random dot noise. Linear detrending is performed on the one-dimensional grayscale signal. Specifically, a linear straight line is fitted to the one-dimensional grayscale signal, which slowly increases or decreases with the vertical position. This straight line is then subtracted from the original one-dimensional grayscale signal. Linear detrending is used to eliminate slow changes in backlight, gradual changes in film thickness, and slow drift in the overall transmittance of the heat insulation film. A linear straight line is used instead of a higher-order curve because a higher-order trend curve may remove the true long-period texture changes. After robust scale normalization, a Hanning window is applied to both ends of the detrended signal, causing the signal to gradually decay at both ends. A fast Fourier transform is then used to convert the one-dimensional grayscale signal into a linear detrending signal. The brightness signal is converted from the spatial domain to the spatial frequency domain. The spatial frequency represents the number of texture cycles that occur per millimeter of the heat insulation film. Local spectral peaks are searched within the effective frequency range. The lower limit of the effective frequency is set to the frequency corresponding to at least 3 complete cycles in the analysis window to exclude slow changes in overall brightness. The upper limit of the effective frequency is set to the frequency corresponding to at least 4 pixels in a complete texture cycle to exclude pixel-level high-frequency noise. The maximum spectral peak is determined. Among all local peaks with an amplitude not less than 60% of the maximum peak, the local peak with the lowest frequency is selected as the fundamental frequency of the good product window. Normal heat insulation film texture is not an ideal sine wave signal. The amplitude of the second harmonic may sometimes be close to or even higher than the fundamental frequency amplitude. If the maximum peak value is selected directly, the second harmonic may be mistakenly identified as the normal main frequency. The median of the fundamental frequencies corresponding to the 20 good product image windows is determined as the normal texture frequency of the current batch, and its reciprocal is determined as the normal average texture period. It should also be noted that after obtaining the normal texture frequency, the dispersion of the frequency of the 20 good product image windows is calculated. When the dispersion of the frequency relative to the median exceeds 15%, it indicates that the texture of the current batch of good products is unstable, or that abnormal areas are mixed in the selected good product window. At this time, the good product window is reselected for calibration. If it still exceeds 15% after reselection, an abnormal process status prompt is output for the current batch, prompting to check the particle dispersion status, stretching parameters, or coating stability.

[0027] S3. Expand the candidate region of the blemishes, extract the one-dimensional grayscale signal of the candidate region of the blemishes and quantify the degree of periodic damage; The degree of periodic disruption includes the amplitude of the dominant frequency and high-frequency noise characteristics; S3.1 For each candidate region of a pattern, the spectrum analysis window is extended forward and backward along the stretching direction of the heat insulation film, with the center position of the candidate region as the reference. The extended spectrum analysis window covers 20 normal texture cycles along the stretching direction. If the length of the candidate region itself exceeds 20 normal texture cycles, the spectrum analysis window completely covers the candidate region, and an additional range of no less than two normal texture cycles is reserved before and after the candidate region. The pixel length of the analysis window is preferably set to an integer power of 2, such as 256, 512 or 1024 pixels, to improve the execution efficiency of the fast Fourier transform. The spectrum analysis window cannot extend to the roll joint, the outside of the film body or invalid image lines. When the candidate region is close to the start position, end position of the roll or invalid image area, the analysis window is shortened according to the actual available image length, and the frequency resolution is re-determined according to the shortened window length. In the horizontal direction, based on the boundary of the candidate region, the width of the candidate region is extended to the left and right by one-quarter. The pixel grayscale of the same vertical position in the horizontal range is averaged to obtain the one-dimensional grayscale signal of the candidate region. S3.2. For the one-dimensional grayscale signal of the candidate region, linear detrending, robust scale normalization, and Hanning window processing are performed sequentially. The specific processing method is consistent with that of the good product one-dimensional signal to ensure the comparability of the spectral characteristics of the good product and the candidate region. A Fast Fourier Transform is performed on the processed one-dimensional signal, and a normal main frequency band is defined based on the normal main frequency. The main frequency band does not contain only one frequency point, but rather frequency tolerances are set on both sides of the normal main frequency. The frequency tolerance is the larger of two frequency resolution units and 8% of the normal main frequency. The two frequency resolution units are used to compensate for frequency dispersion caused by the finite window, and the 8% normal main frequency tolerance is used to adapt to slight changes in stretching speed and particle spacing during normal production. A second harmonic band is set near the position of twice the normal structural frequency. The second harmonic band is retained as a normal structural frequency and is not included in the defect noise. The reason is that the normal heat insulation film texture is usually not an ideal sine curve, and the normal texture may also naturally generate a strong second harmonic. The dominant frequency maintenance amplitude of the candidate region is calculated. The dominant frequency maintenance amplitude is used to represent how much of the effective spectral energy of the candidate region is still concentrated near the normal dominant frequency. That is, the proportion of energy in the effective spectral energy located in the normal dominant frequency band. When the heat insulation film particles are arranged normally, the spectral energy is mainly concentrated near the normal dominant frequency, and the dominant frequency maintenance amplitude is high. When the particles are clustered, missing, or the spacing is disordered, the original period is destroyed, the proportion of energy near the dominant frequency decreases, and the dominant frequency maintenance amplitude decreases accordingly. S3.3 Further calculate the high-frequency diffusion energy and non-dominant frequency diffusion energy of the candidate region. The high-frequency diffusion energy is the proportion of spectral energy that is 1.5 times higher than the normal dominant frequency and not located within the second harmonic band. This reflects high-frequency anomalies formed by particle aggregation edges, particle void boundaries, abrupt changes in local spacing, and fine textures. The non-dominant frequency diffusion energy is the proportion of spectral energy outside the dominant frequency band and second harmonic band within the effective frequency range (not higher than 1.5 times the normal dominant frequency). This reflects spectral diffusion caused by large-area particle absence, irregular spacing, and slow density changes. The diffuse energy and non-dominant frequency diffuse energy are weighted and fused into high-frequency noise features. The preferred weighting values ​​are 0.7 and 0.3. The high-frequency diffuse weight is set to 0.7 because the patterns on the heat insulation film are usually accompanied by local particle aggregation, void boundaries and abrupt changes in particle spacing, and the high-frequency anomalies are more obvious. Retaining the non-dominant frequency diffuse weight of 0.3 can avoid missing light spots with large areas but slow gray-scale changes. The amplitude of the dominant frequency and the high-frequency noise features are combined to form two-dimensional spectral feature points. At the same time, the complete spectral curve of the candidate region, the peak position of the dominant frequency, the analysis window range and the texture anomaly score are preserved.

[0028] S4. Establish batch adaptive elliptical decision boundaries and assign weights to candidate areas of the blemishes. For each assigned blemish area, refine the pixel boundaries of the blemish defects and calculate the physical area and visual deviation. Further calculate the blemish visibility index and output graded alarms and roll material location. S4.1 For each type of heat insulation film, establish a model-level normal spectral feature library. Specifically, pre-collect no fewer than 500 manually confirmed good product image windows and calculate the dominant frequency holding amplitude and high-frequency noise characteristics of each good product window to establish a model-level normal spectral feature library. Based on the model-level normal spectral feature library, calculate the normal spectral feature center and the normal spectral feature covariance matrix. The normal spectral feature center is used to represent the typical position of the good product of this model in the two-dimensional space of "dominant frequency holding amplitude - high-frequency noise characteristics". The covariance matrix is ​​used to describe the fluctuation degree of the two spectral features and the correlation between them. The model-level normal spectrum model is adaptively adjusted using the spectral characteristics of the current batch of good products. The batch adaptive weight is calculated by the number of samples of the current batch of good products and the equivalent number of samples of the historical model. Specifically, the batch adaptive weight is obtained by dividing the number of samples of the current batch of good products by the sum of the number of samples of the current batch of good products and the equivalent number of samples of the historical model. The adjusted normal feature center is obtained by weighting the model-level historical feature center and the current batch of good product feature center. The proportion of the current batch is the batch adaptive weight, and the sum of the proportion of the model-level historical model and the proportion of the current batch is 1. The normal covariance matrix is ​​adjusted by weighting in the same way. The number of equivalent samples for the historical model is determined based on the dispersion of feature centers among the most recent qualified batches. The greater the difference between batches, the smaller the number of historical equivalent samples, which increases the weight of the current batch. The smaller the difference between batches, the larger the number of historical equivalent samples, which keeps the historical model stable. Perform candidate screening on the complete image of good products at the model level, and retain only the regions that were mistakenly screened as candidate regions but were confirmed to be normal by manual inspection to form a normal candidate feature library. Use this normal candidate feature library and the normal candidate regions of the current batch to jointly establish an elliptical decision boundary. The adjusted covariance matrix is ​​subjected to shrinkage regularization. Covariance shrinkage is used to prevent the covariance matrix from approaching singularity when the number of samples in the current batch is small, which would cause the ellipse boundary to be too narrow, too long, or have an abnormal orientation. If the shrinkage coefficient is too large, the ellipse will approach a circle and cannot accurately reflect the correlation between the main frequency amplitude and the high-frequency noise characteristics. If no shrinkage is performed, the model stability is poor. For the candidate region of the pattern, the ellipse distance is calculated based on the adjusted normal feature center and the shrunken covariance matrix. The corresponding ellipse distance is calculated using the model-level good sample, and the 99.5 percentile of the good sample ellipse distance is determined as the pattern weighting threshold. When the ellipse distance of the candidate region is less than or equal to the pattern weighting threshold, it is judged as normal texture fluctuation, indicating that although there is a certain texture abnormality in the region, its periodic spectrum characteristics are still within the normal joint fluctuation range. If the ellipse distance of the candidate region is greater than the pattern weighting threshold, it is judged as a real pattern defect region, indicating that the normal texture main frequency has been significantly weakened or the non-main frequency energy has been significantly increased. Specifically, elliptical distance Represented as:

[0029] in It is a two-dimensional spectral feature. For normal feature centers, It is the covariance matrix; For multiple weighted regions that overlap spatially, are directly adjacent, or are spaced apart by less than one normal texture cycle, they are merged into the same mottled defect object. If the interval between two weighted regions is greater than one normal texture cycle, and the texture and spectrum of the middle region have been restored to normal, they are retained as two independent mottled defects. The elliptical distance, dominant frequency preservation amplitude, high-frequency noise characteristics, texture anomaly score, image position, and physical position of each weighted mottled defect are saved simultaneously.

[0030] S4.2 For each confirmed mottled area, a local background ring is established around it. The background ring is set outside the confirmed area, and its width is preferably one-quarter of the side length of the image block. When selecting background ring pixels, other confirmed mottled areas, membrane outer areas, membrane edges, invalid image lines and obvious contamination points are excluded to prevent abnormal pixels from participating in normal background estimation. When the number of effective background pixels is less than 60% of the total number of background ring pixels, the range of the background ring is gradually expanded until a sufficient number of normal background pixels are obtained. If a sufficient number of background pixels cannot be obtained after expansion, the defect is marked as an edge region defect and supplemented using the background parameters of the nearest normal image block. A local grayscale plane is fitted using grayscale pixels within the background ring. The local grayscale plane represents the normal grayscale that each position in the designated area should have in the absence of blemishes. For each pixel in the designated area, the absolute difference between its actual corrected grayscale and the corresponding local background grayscale is calculated. Then, based on the residual between the actual grayscale and the fitted background grayscale within the background ring, the background robust fluctuation scale is calculated. Pixels with a grayscale difference exceeding three times the background robust fluctuation scale are marked as initial blemish pixels. The use of a three-times background fluctuation scale is to keep most normal grain textures and random imaging noise within the normal range, while pixels that deviate significantly from the local background are classified as defect areas. The initial blemish pixels undergo connected component processing. Specifically, isolated regions with an area less than one-tenth of the minimum detectable blemish area are deleted to exclude camera defects, dust, or single abnormal particles. Small holes formed by a few normal pixels are filled within the blemish body. For narrow regions connected by only one or two pixels, merging is determined based on the continuity of grayscale on both sides. The pixel region connected to the center of the confirmed region is identified as the final blemish mask. When two unconnected blemish pixel regions exist within a confirmed region, and both meet the minimum defect area requirement, they are split into two independent blemish defect objects. The number of pixels contained in the final blemish mask is counted. Converting this to actual physical area, for each blemish pixel, calculate the absolute difference between its actual grayscale and the local background grayscale, then divide by the corresponding local background grayscale to obtain the relative grayscale deviation of that pixel. Average the relative grayscale deviations of all blemish pixels to obtain the average relative grayscale deviation of the blemish defect. Simultaneously, calculate the longitudinal length, transverse width, aspect ratio, outline perimeter, and distance from the left and right edges of the membrane. The longitudinal length reflects the extent of the blemish extension along the stretching direction, the transverse width reflects the range of influence of the blemish on the width direction of the membrane, and the distance from the membrane edge is used to help distinguish edge coating fluctuations from blemishes in the middle of the membrane.

[0031] S4.3 For each blemish defect, calculate the blemish visibility index. Specifically, it is obtained by multiplying the actual physical area of ​​the blemish defect by the average relative grayscale deviation. The product of the defect area and the average relative grayscale deviation is used because the visual visibility of the blemish is affected by both the range and the visual contrast. Small spots with obvious grayscale differences, as well as large spots with weak grayscale differences, may both reach the level of visual visibility. Using only the area threshold or the grayscale threshold cannot stably cover different types of blemishes. To determine the acceptance threshold for each type of heat insulation film, specifically, for each film model, calibration samples with different areas and grayscale deviations are pre-collected. Under uniform lighting conditions, observation distance, and observation time, at least three inspectors independently determine whether the patterns in the samples are visually visible. The majority opinion is used as the final visual label for the sample. Different visibility index thresholds are then traversed, and the false alarm rates for visually visible patterns judged as permissible texture fluctuations and visually invisible textures judged as unacceptable patterns are calculated. The visibility index corresponding to the lowest weighted false alarm cost and weighted false alarm cost is selected as the acceptance threshold for that type of heat insulation film. The preferred weight for the false alarm cost is 4, and the preferred weight for the false alarm cost is 1. That is, the cost of missing a visually visible pattern is calculated as four times that of falsely reporting a normal texture area. A missed detection may lead to... Defective heat insulation film enters the winding, slitting, or delivery stages. False alarms usually only increase the workload of manual re-inspection. Therefore, it is necessary to set a higher weight for missed detections. The defect level of the heat insulation film is judged based on the pattern visibility index and the acceptance threshold. If the pattern visibility index is less than or equal to the acceptance threshold, it is judged as Level 0, which allows for texture fluctuations. Although there are quantifiable micro-uniformities in this area, they have not yet reached the level of visual non-compliance. No production alarm is triggered, and only the detection data is saved. If the pattern visibility index is greater than the acceptance threshold but less than or equal to twice the acceptance threshold, it is judged as Level 1, which is visually visible. The pattern position is marked in the detection interface, triggering an audio-visual prompt and generating a local defect image and outline for the operator to review. When the pattern visibility index is greater than twice the acceptance threshold, it is judged as Level 2, which is obviously non-compliant. A high-level alarm is output, and a coding mark, labeling, speed reduction, or shutdown signal is sent according to the production line control requirements. The reason for using twice the visibility threshold as the judgment line for obviously unacceptable marks is that when the visibility index reaches twice the acceptance boundary, under the condition that other conditions remain unchanged, it means that the defect area, visual contrast or the combination of the two has obviously exceeded the allowable range, and has a high certainty of being unacceptable. Based on the mapping relationship between the original image and the coordinates of the heat insulation film, the longitudinal distance from the start end of the roll of each blemish defect to the lateral distance from the left edge of the heat insulation film, as well as the longitudinal and lateral start and end positions of the defect, are calculated. For each blemish, defect data is output, including heat insulation film model, production batch, roll number, acquisition time, defect level, local defect image, pixel-level defect outline, longitudinal position, lateral position, defect area, average relative grayscale deviation, visibility index, main frequency holding amplitude, high-frequency noise characteristics, ellipse distance, and texture anomaly score. For the same type of blemish that appears continuously in the same roll of heat insulation film, trend statistics are performed according to defect location and occurrence frequency. When multiple level 2 blemishes appear consecutively within a short longitudinal distance, or when the number of level 1 blemishes per unit length continues to increase, a process trend warning is output, including prompts to check the particle dispersion status, coating uniformity, production line stretching speed, curing temperature, or roller surface cleaning status. After a roll of heat insulation film is completed, the normal areas and blemish defect areas that have been manually verified are saved to the corresponding model's sample library.

[0032] This embodiment also provides an image-based thermal insulation film defect identification system, including: The image processing module is used to establish the polarization imaging environment of the heat insulation film and acquire the original image with displacement synchronization, perform imaging correction on the original polarization image and establish the physical coordinate image of the heat insulation film. The benchmark analysis module is used to perform directional texture analysis on standardized heat insulation film images and screen candidate areas for patterns, collect good product images of each type of heat insulation film, and establish a normal cycle benchmark for heat insulation film using the good product images of the current batch. The candidate region quantization module is used to expand the mottled candidate region, extract the one-dimensional grayscale signal of the mottled candidate region, and quantify the degree of periodic damage. The mottled pattern recognition output module is used to establish batch adaptive elliptical decision boundaries and assign weights to candidate mottled patterns. For each assigned mottled pattern region, it refines the pixel boundaries of the mottled defects and calculates the physical area and visual deviation degree. It further calculates the mottled pattern visibility index and outputs graded alarms and roll material location. In summary, this invention achieves stable correspondence between image rows and the actual position of the membrane through displacement-synchronous polarization acquisition and dark field and uniform field correction, while suppressing reflection, uneven illumination, and pixel response differences. It reduces the number of normal areas entering spectral analysis through initial screening with a directional gray-level co-occurrence matrix, improving online processing efficiency. It quantifies the disruption of the periodic arrangement of functional particles through batch main frequency calibration, main frequency amplitude preservation, and non-main frequency dispersion energy extraction. It enhances the adaptability of judgment under different batch conditions by jointly constructing elliptical boundaries using historical model and current batch features. It improves the accuracy of positioning and size evaluation by refining defect boundaries and converting physical area through local background fitting. Finally, it enables the detection results to serve re-inspection, coding, slitting, and production alarms through visibility index, visual calibration threshold, and roll material coordinate output, thereby reducing excessive rejection caused by micro-texture fluctuations.

[0033] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for identifying defects in heat insulation films based on images, characterized in that: include, Establish a polarization imaging environment for the heat insulation film and acquire original images with displacement synchronization. Perform imaging correction on the original polarization images and establish a physical coordinate image of the heat insulation film. Directional texture analysis was performed on standardized heat insulation film images to screen candidate areas for patterns. Good product images of each type of heat insulation film were collected, and a normal cycle benchmark for heat insulation film was established using the good product images of the current batch. The candidate regions for blemishes are expanded, and the one-dimensional grayscale signals of the candidate regions are extracted and the degree of periodic damage is quantified. The degree of periodic disruption includes the amplitude of the main frequency maintenance and high-frequency noise characteristics; Establish batch adaptive elliptical decision boundaries and assign weights to candidate areas of the blemishes. For each assigned blemish area, refine the pixel boundaries of the blemish defects and calculate the physical area and visual deviation. Further calculate the blemish visibility index and output graded alarms and roll material location.

2. The image-based heat insulation film defect identification method as described in claim 1, characterized in that: The specific steps for establishing a polarization imaging environment for the heat insulation film and acquiring original images with synchronous displacement are as follows: a high-definition line array camera is vertically installed above the heat insulation film production line, so that the pixel arrangement direction of the line array camera is consistent with the width direction of the heat insulation film. When the heat insulation film moves continuously along the production line, the line array camera acquires images line by line, and each acquired line is stitched together in sequence according to the movement order of the film body to form a two-dimensional original image that simultaneously covers the longitudinal and transverse positions of the heat insulation film. A uniform backlight is set on the other side of the heat insulation film corresponding to the line scan camera, so that the backlight passes through the heat insulation film and enters the line scan camera. A first polarizer is set on the light-emitting surface of the backlight, and a second polarizer is set in front of the camera lens. The transmission directions of the two polarizers are set to be approximately orthogonal, and the included angle between the two polarizers is calibrated. A rotary encoder is installed on the production line drive roller or guide roller corresponding to the line scan camera and connected to the line scan camera to define the physical size of the longitudinal pixels and the physical size of the transverse pixels of the heat insulation film.

3. The image-based heat insulation film defect identification method as described in claim 2, characterized in that: The specific steps for performing imaging correction on the original polarization image and establishing the physical coordinate image of the heat insulation film are as follows: Before formal production testing, the backlight is turned off, and the gain, exposure time, and acquisition frequency of the linear array camera are kept consistent with the formal testing state. 100 lines of dark field images are continuously acquired. For each pixel position in the linear array camera, the average value of the dark field grayscale is calculated 100 times to obtain the dark field reference for that pixel position. Then, the backlight is turned on, and a good quality heat insulation film that has been manually confirmed to be free of spots and has uniform light transmission is placed at the testing position. 100 lines of uniform field images are continuously acquired. The average grayscale is calculated for each linear array pixel position to obtain the uniform field response for that pixel position. For each original pixel acquired during formal production, the dark field reference corresponding to the same linear array position is subtracted first, and then divided by the difference between the uniform field response and the dark field reference at that position to obtain the normalized corrected grayscale. Calculation results less than 0 are set to 0, and calculation results greater than 1 are set to 1, so that the corrected grayscale is within the range of 0 to 1. By utilizing the grayscale difference between the membrane area and the background of the field of view, the left and right boundaries of the heat insulation membrane are found in each image row, and the background area outside the boundaries is excluded. The encoder's cumulative position when a roll of heat insulation film enters the inspection station is defined as the longitudinal starting point, and the lateral position corresponding to the left edge of the heat insulation film is defined as the lateral starting point. Based on the longitudinal and lateral starting points, the original image pixels are mapped to the heat insulation film coordinates to form a standardized heat insulation film image.

4. The image-based defect identification method for heat insulation films as described in claim 3, characterized in that: The specific steps for performing directional texture analysis and screening candidate regions for pattern patterns on standardized heat insulation film images are as follows: the standardized heat insulation film image is divided into overlapping rectangular image blocks. Local gray-level normalization is performed on each image block, the median gray level of the image block is calculated, and the median absolute deviation of each pixel gray level relative to the median is calculated. The median absolute deviation is used to represent the local gray-level fluctuation scale of the image block. The locally normalized gray level is quantified into 32 levels. Gray-level co-occurrence matrices are established along the stretching direction and the width direction of the heat insulation film, respectively. Two pixel distances are used in each direction. For each direction and each pixel distance, four texture statistics, namely energy, entropy, contrast and correlation, are extracted from the gray-level co-occurrence matrix. Weighted fusion is performed on texture amounts of the same type at a distance of two pixels, and further weighted fusion is performed on texture amounts in the stretching direction and the film width direction; For each type of heat insulation film, good product image blocks are collected in advance, and the normal median and normal fluctuation range of its energy, entropy, contrast and correlation are statistically analyzed. For the image patch to be detected, the degree of decrease in energy relative to the good benchmark, the degree of increase in entropy relative to the good benchmark, the degree of increase in contrast, and the degree of decrease in correlation are calculated and a texture anomaly score is constructed. The 99th percentile of the texture anomaly score of the good image patch is determined as the candidate screening threshold. When the texture anomaly score of the image patch to be detected exceeds the candidate screening threshold, it is marked as a pattern candidate patch; otherwise, it is identified as a normal texture image patch. Candidate image patches that overlap spatially or are connected by eight neighboring regions are merged into the same pattern candidate region, and the vertical start and end positions, horizontal start and end positions, center position, coverage area, and maximum texture anomaly score of the candidate region are recorded.

5. The image-based heat insulation film defect identification method as described in claim 4, characterized in that: The specific steps for establishing a normal cycle benchmark for heat insulation film using the images of good products from the current batch are as follows: After each batch of heat insulation film starts production and enters a stable operating state, select good product image windows from the images of the current batch that have been manually confirmed to be free of blemishes. Each good product image window covers 20 normal texture cycles along the stretching direction. For each good product image window, average the pixel grayscale at the same longitudinal position along the film width direction to obtain a one-dimensional grayscale signal that varies along the stretching direction of the heat insulation film. Linear detrending processing is performed on the one-dimensional grayscale signal. Hanning windows are applied to both ends of the detrended signal. The one-dimensional grayscale signal is converted from the spatial domain to the spatial frequency domain using fast Fourier transform. Local spectral peaks are found within the effective frequency range. The maximum spectral peak is determined. Among all local peaks with an amplitude not less than 60% of the maximum peak, the local peak with the lowest frequency is selected as the fundamental frequency of the good window. The median of the base frequencies corresponding to the 20 good product image windows is determined as the normal texture frequency of the current batch, and its reciprocal is determined as the normal average texture period.

6. The image-based defect identification method for heat insulation films as described in claim 5, characterized in that: The extraction of one-dimensional grayscale signals of candidate regions for blemishes and quantification of periodic damage are specifically as follows: For each candidate region for blemishes, with the center position of the candidate region as a reference, the spectrum analysis window is extended forward and backward along the stretching direction of the heat insulation film. The extended spectrum analysis window covers 20 normal texture periods along the stretching direction. In the horizontal direction, based on the boundary of the candidate region, the width of the candidate region is extended to the left and right sides by one-quarter. The pixel grayscale of the same vertical position in the horizontal range is averaged to obtain the one-dimensional grayscale signal of the candidate region. Linear detrending, robust scaling and Hanning window processing are performed sequentially on the one-dimensional grayscale signal of the candidate region. Fast Fourier Transform is performed on the processed one-dimensional signal. The normal main frequency band is defined according to the normal main frequency. The second harmonic band is set near twice the normal main frequency. The main frequency holding amplitude of the candidate region is calculated. The high-frequency diffusion energy and non-dominant frequency diffusion energy of the candidate region are further calculated. The high-frequency diffusion energy and non-dominant frequency diffusion energy are weighted and fused into high-frequency noise features. The dominant frequency amplitude and high-frequency noise features are combined to form two-dimensional spectral feature points. At the same time, the complete spectral curve, dominant frequency peak position, analysis window range and texture anomaly score of the candidate region are saved.

7. The image-based heat insulation film defect identification method as described in claim 6, characterized in that: The specific steps for establishing batch adaptive elliptical decision boundaries and determining the weights of candidate regions for patterns are as follows: For each type of heat insulation film, a model-level normal spectrum feature library is established, and the normal spectrum feature center and normal spectrum feature covariance matrix are calculated based on the model-level normal spectrum feature library. The normal spectrum model at the model level is adaptively adjusted using the spectral characteristics of the current batch of good products. The adjusted normal feature center is obtained by weighting the historical feature center at the model level and the feature center of the current batch of good products. The adjusted covariance matrix is ​​shrunk and regularized. For the candidate region of the blemish, the elliptical distance is calculated based on the adjusted normal feature center and the shrunk covariance matrix. The corresponding elliptical distance is calculated using the model-level good product samples. The 99.5 percentile of the good product elliptical distance is determined as the blemish weighting threshold. When the elliptical distance of the candidate region is less than or equal to the blemish weighting threshold, it is judged as normal texture fluctuation. If the elliptical distance of the candidate region is greater than the blemish weighting threshold, it is judged as a real blemish defect region. For multiple confirmed regions that overlap spatially, are directly adjacent, or are spaced apart by less than one normal texture cycle, they are merged into the same pattern defect object. If the interval between two confirmed regions is greater than one normal texture cycle, and the texture and spectrum of the middle region have been restored to normal, they are retained as two independent pattern defects.

8. The image-based defect identification method for heat insulation films as described in claim 7, characterized in that: The process of refining the pixel boundaries of the blemish defects and calculating the physical area and visual deviation is as follows: For each confirmed blemish region, a local background ring is established around it. When the number of effective background pixels is less than 60% of the total number of pixels in the background ring, the range of the background ring is gradually expanded until a sufficient number of normal background pixels are obtained. If a sufficient number of background pixels cannot be obtained after expansion, the defect is marked as an edge region defect and supplemented with the background parameters of the nearest normal image block. A local grayscale plane is fitted using grayscale pixels within the background ring band. The local grayscale plane is used to represent the normal grayscale that each position in the confirmation area should have when there are no spots. For each pixel in the confirmation area, the absolute difference between its actual corrected grayscale and the corresponding local background grayscale is calculated. Then, based on the residual between the actual grayscale in the background ring band and the fitted background grayscale, the background robust fluctuation scale is calculated. Pixels with a grayscale difference exceeding 3 times the background robust fluctuation scale are marked as initial spot pixels. Connectivity processing is performed on the initial blemish pixels, and the pixel region connected to the center of the confirmation region is determined as the final blemish mask. The number of pixels contained in the final blemish mask is counted and converted into the actual physical area. For each blemish pixel, the absolute difference between its actual gray level and the local background gray level is calculated, and then divided by the corresponding local background gray level to obtain the relative gray level deviation of the pixel. The average relative gray level deviation of all blemish pixels is calculated to obtain the average relative gray level deviation of the blemish defect. At the same time, the longitudinal length, transverse width, aspect ratio, contour perimeter, and distance from the left and right edges of the membrane are calculated for the blemish defect.

9. The image-based defect identification method for heat insulation films as described in claim 8, characterized in that: The specific steps for calculating the mottle visibility index and outputting graded alarms and roll material location are as follows: For each blemish defect, calculate the blemish visibility index; Obtain the acceptance threshold for each type of heat insulation film. Determine the defect level of the heat insulation film based on the pattern visibility index and the acceptance threshold. If the pattern visibility index is less than or equal to the acceptance threshold, it is judged as level 0, allowing texture fluctuations. No production alarm is triggered, and only the test data is saved. If the pattern visibility index is greater than the acceptance threshold but less than or equal to twice the acceptance threshold, it is judged as level 1, visually visible pattern. Mark the pattern position in the test interface, trigger an audio-visual prompt, and generate a defect local image and outline for the operator to review. When the pattern visibility index is greater than twice the acceptance threshold, it is judged as level 2, obviously unqualified pattern. A high-level alarm is output, and a coding mark, labeling, speed reduction, or shutdown signal is sent according to the production line control requirements. Based on the mapping relationship between the original image and the coordinates of the heat insulation film, the longitudinal distance from the start end of the roll of each blemish defect, the lateral distance from the left edge of the heat insulation film, and the longitudinal and lateral start and end positions of the defect are calculated. Defect data is output for each blemish. For the same type of blemish that appears continuously in the same roll of heat insulation film, trend statistics are performed according to the defect location and frequency of occurrence. When multiple level 2 blemishes appear continuously within a short longitudinal distance, or when the number of level 1 blemishes per unit length continues to increase, a process trend warning is output. After a roll of heat insulation film is completed, the normal areas and blemish defect areas that have been manually verified are saved to the sample library of the corresponding model.

10. An image-based thermal insulation film defect identification system, based on the image-based thermal insulation film defect identification method according to any one of claims 1 to 9, characterized in that: include, The image processing module is used to establish the polarization imaging environment of the heat insulation film and acquire the original image with displacement synchronization, perform imaging correction on the original polarization image and establish the physical coordinate image of the heat insulation film. The benchmark analysis module is used to perform directional texture analysis on standardized heat insulation film images and screen candidate areas for patterns, collect good product images of each type of heat insulation film, and establish a normal cycle benchmark for heat insulation film using the good product images of the current batch. The candidate region quantization module is used to expand the mottled candidate region, extract the one-dimensional grayscale signal of the mottled candidate region, and quantify the degree of periodic damage. The pattern recognition output module is used to establish batch adaptive elliptical decision boundaries and assign weights to candidate pattern regions. For each assigned pattern region, the pixel boundaries of the pattern defect are refined and the physical area and visual deviation are calculated. The pattern visibility index is further calculated and the graded alarm and roll material location are output.

Images